Player is loading...
Q&A
Dr. Dinesh Babu Jayagopi, Assistant Professor at IIIT Bangalore
Thursday, June 13, 2019 · 12:03 p.m. · 09m 25s
Embed
Transcriptions
Note: this content has been automatically generated.
00:00:07
e. a.
00:00:13
if you right
00:00:18
or e. e. i. e.
00:00:23
oh yeah so it does it does like that only so maybe ten fold the
00:00:30
but only
00:00:34
oh oh i use e. e. so so i think that there are balanced right so there
00:00:49
are uh the data was balanced uh so that was not a you should it's o. e.
00:00:59
yeah oh whoa o. e. r. e.
00:01:08
so of course i mean one of one of validation came from actually implying it
00:01:13
in the in the wild sort of setting i mentioned do you like to there
00:01:18
somebody who's external ah started using these ratings
00:01:22
and uh they did have their own rating alongside
00:01:26
and they trusted that yeah this is this is this is something that
00:01:31
ah that's maybe mistakes which are like one point uh here and there but pretty much on i'll
00:01:37
similar to what the human ratings read well so so that is one validation
00:01:43
uh we do have 'em up of course yeah i mean having larger numbers would be the uh
00:01:50
would be the base rate to go in terms of statistics
00:01:54
but yeah this is another direction uh possibly a validating the more
00:02:00
yeah
00:02:06
um hum
00:02:10
uh_huh uh_huh oh
00:02:14
oh ah he's a s. t. ah so
00:02:24
oh was oh so the follow up question generation it was a it was on the data set
00:02:39
so this was on the text data set of
00:02:44
right so so if you remember we had the the ten interview so it was it was based on that yes
00:02:51
and uh e. m. i. a. o.
00:02:56
so she using a model which needs a which needs a data set
00:03:02
as well as it needs these ah say shock sentences and focused tokens right um
00:03:11
so um so i mean so these data sets so s.
00:03:15
uh the data for the cute in it a mortal yes
00:03:21
oh
00:03:23
oh
00:03:28
yes
00:03:35
oh
00:03:37
uh_huh uh yes i
00:03:42
yes yes ah ah ah no no no so these are the
00:03:50
uh as you are likely saying e. these uh uh where you say
00:03:54
fairly general data sets i'm is on a data set is about their products and so on
00:04:01
it's quite is uh uh is the design of the data that is commonly used for other task so i so that
00:04:07
is one limitation as in i'll be do have one data
00:04:11
set bad we have this q. a. and a human generated ah
00:04:16
follow up question say the follow up questions and it'll be human but we haven't patch
00:04:21
that uh it hasn't it would be good to ah sort of stanton this followup question
00:04:27
um maybe be is dawn um based on this data say that you have that is one possibility um
00:04:35
i i mean also i mean again as i was telling you um in terms of the structure nature and so
00:04:41
on that uh i mean the the set of questions which are um which i asked in these hatch our interviews are
00:04:48
uh i'm having a second nature right uh so how you bring in that uh nature maybe as a knowledge
00:04:55
graph or uh some ah prior knowledge or something of that there is also an open uh open question yes
00:05:15
yes
00:05:19
so i guess or recognition yeah so
00:05:22
so we did well i mean badly we also are working on a guy said recognition but in terms of
00:05:29
in terms of what that is that a lot of work on a guy said recognition
00:05:33
uh but there is limited work on a guest a synthesis so that way there's
00:05:39
sort of a a receipt motivated to look at that from that point of view
00:05:44
but we also uh i mean from the point of
00:05:46
view of transcribing um existing videos where somebody's a signing
00:05:52
ah so that you can get some annotations you can get some uh
00:05:56
sort of a transcriptions using gets a recognition from that point of view
00:06:01
but to help this problem we're working on a guess that recognition as well yes
00:06:06
but it is yeah fairly studied uh the number of work on synthesis less
00:06:14
thanks
00:06:18
or should we just convinced everyone see exactly this question yes
00:06:30
so then another set of hundred questions uh from which a five question the sample
00:06:36
but we also sorry i think i i might uh i made one several so
00:06:41
the set this common but the questions are not come so though five
00:06:46
question save for the return after the spoken would be different from the five
00:06:50
question that they uh on certain the spoken so it would be sample sites
00:06:55
that if the ability on so those we will or to some other white
00:07:00
so it is exactly not the scene but it is in the set of say
00:07:05
questions which are like usually asked in talking to you from that sex is under
00:07:11
t. yes uh_huh
00:07:19
f. has it
00:07:23
e. s. yes uh_huh and do the question is in use
00:07:29
that as a yes okay to say yeah yes yes yes yes
00:07:37
okay it's a show and e. uh yeah
00:07:45
in a shoe is it also the hum hum
00:07:51
yes i uh so the so the feedback comes from a few rating
00:07:57
a set of feedbacks from a trainers who uh on the job
00:08:03
so this is a decision that we took a so we will uh
00:08:06
we will ask people how war uh the actionable feedbacks you will you uh
00:08:13
in general for people and we also look at some of these views uh that uh
00:08:19
we had a back then so and then now come up with a list and then
00:08:24
these are the sort of the consensus a consensus questions
00:08:28
across multiple uh trainers on the job types use the expert
00:08:37
yes yes yes and you know the shoes
00:08:42
yes and and then these people also uh i mean how much from that not the set
00:08:46
a couple of them also gave a ground truth on the data set ah that ah
00:08:52
that we had collected so for every person they would say it is on or off
00:08:56
ah uh these uh feedback so we have that don't shoot
00:09:00
uh data and then the uh model would learn and on the test
00:09:06
set it would try to predict uh which one is on off ah
00:09:12
yeah
00:09:14
uh_huh
Conference Program
Idiap Speaker Series: Investigating Multiple facets of communication skill assessment and feedback
Dr. Dinesh Babu Jayagopi, Assistant Professor at IIIT Bangalore
June 13, 2019 · 11:06 a.m.
231 views
